Claude Sonnet 5 Basically Closed the Opus Gap

Sonnet has always been the cheap workhorse. Good enough, not great, the model you reached for when Opus felt like overkill.

Sonnet 5 just moved the goalposts on what "good enough" means.

It lands near Opus 4.8 on coding and reasoning, and right now it runs at $2 per million input tokens instead of $5. That's not a minor discount, that's Anthropic handing you flagship-adjacent performance at less than half the price.

The Part That Actually Got My Attention

A tester asked it to investigate a bug. Nobody told it to write a test. Nobody told it to verify anything.

Unprompted, it wrote a test to reproduce the bug, fixed the code, then stashed its own fix just to confirm the bug came back without it. One pass, no hand-holding.

That's not autocomplete with good taste. That's a model checking its own work because it assumed you'd want proof, not just a claim.

The old Sonnet finished tasks. This one finishes the doubt.

The old Sonnet had a habit of quitting halfway through multi-step agentic work. You'd hand it something with five moving parts and get three back, along with a polite explanation of why the other two were tricky. Sonnet 5 doesn't do that. It follows the work through to the end, and it seems to actually care whether the end is correct.

Why This Matters for Your Workflow

If you were routing anything to Opus purely because Sonnet couldn't be trusted to finish or self-verify, that calculation just changed. Less than half the price, near-flagship results, and a model that audits itself before handing you the output.

Go test it on whatever job you were previously too nervous to give to Sonnet. That's the actual signal here, not the benchmark numbers, the fact that it stopped needing supervision to know when it was wrong.

#ClaudeSonnet5 #Anthropic #AInews #AIWorkflows